This article was downloaded by: [157.182.150.

22] On: 08 June 2014, At: 12:47

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Interfaces
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

On the Development of a Soccer Player Performance

Rating System for the English Premier League
Ian G. McHale, Philip A. Scarf, David E. Folker,

To cite this article:

Ian G. McHale, Philip A. Scarf, David E. Folker, (2012) On the Development of a Soccer Player Performance Rating System for
the English Premier League. Interfaces 42(4):339-351. http://dx.doi.org/10.1287/inte.1110.0589

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2012, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 Vol. 42, No. 4, July–August 2012, pp. 339–351
ISSN 0092-2102 (print) — ISSN 1526-551X (online) http://dx.doi.org/10.1287/inte.1110.0589
© 2012 INFORMS

On the Development of a Soccer Player

Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.

Performance Rating System for the English

Premier League
Ian G. McHale
Centre for Operations Management, Management Science and Statistics, and Centre for Gambling Studies,
Salford Business School, University of Salford, Salford M5 4WT, United Kingdom, i.mchale@salford.ac.uk

Philip A. Scarf
Centre for Operations Management, Management Science and Statistics,
Salford Business School, University of Salford, Salford M5 4WT, United Kingdom, p.a.scarf@salford.ac.uk

David E. Folker
Football DataCo Ltd., London W1U 8PL, United Kingdom, dfolker@football-dataco.com

The EA Sports Player Performance Index is a rating system for soccer players used in the top two tiers of
soccer in England—the Premier League and the Championship. Its development was a collaboration among
professional soccer leagues, a news media association, and academia. In this paper, we describe the index and
its construction. The novelty of the index lies in its attempts to rate all players using a single score, regardless
of their playing specialty, based on player contributions to winning performances. As one might expect, players
from leading teams lead the index, although surprises happen.
Key words: sports and recreation; football; performance measurement; ranking.
History: This paper was refereed. Published online in Articles in Advance April 6, 2012.

P erformance assessment is a fundamental tool for

quantitative analysts and operational researchers.
For example, fund managers assess the performance
of traders (Grinblatt and Titman 1989), and engi-
neers monitor the performance of production lines
(Swamidass and Newell 1987). In sports, performance
measurement typically takes the form of a rating sys-
tem (i.e., players receive points for performances) or a
rankings list (i.e., players’ performances are ordered).
Many stakeholders in the sports industry, includ-
ing the teams, fans, and pundits, use rating sys-
tems. In baseball, for example, the Oakland Athletics
used a quantitative analyst to value players (Lewis
2003). The coaching staff used the values to iden-
tify and rank the most desirable players in the draft.
By drafting and signing players who were effectively
undervalued by the market, the team was able to
compete with much wealthier teams for a fraction of
the cost. Fans use official ratings to gauge how well
their favourite player or team is performing, and then
339
debate rating and ranking lists among friends or on
social networking websites.
Rating systems already exist in many sports. In-
dividual sports (e.g., tennis) provide a simple frame-
work for rating individual players. The official
world rankings of the Association of Tennis Pro-
fessionals (ATP) are derived from results in tour-
naments from the previous 52 weeks of compe-
tition. Its website (http://www.atpworldtour.com/
Rankings/Singles.aspx) lists the most current rank-
ings. The tournament’s prestige and a player’s
progress through the tournament determines how
points are awarded. Although the ATP world rank-
ings are well established, they have been called
into question in the academic literature. McHale and
Morton (2011) show that the points that underlie these
rankings perform poorly as a forecasting tool. Golf
has a similar story; McHale and Forrest (2005) show
the Official World Golf rankings list to be a relatively
poor predictor of the result of the next tournament. It

340
appears, then, that ranking in individual sports is not
simple.
Rating teams in team sports is similar to rat-
ing individuals in individual sports. It is often a
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
simpler exercise because fewer teams are compet-
ing against each other, and one can organise and
design leagues with the objective of ranking (and
to some extent rating) the teams at the end of
the season. For example, in a round-robin league,
each team plays the other teams in turn and points
are awarded for success. The National Football
League regional leagues are examples of a round-
robin league, and the final league table provides a
list of ratings for the teams in that region. Where
teams do not play in leagues, sports associations
have typically commissioned consultants to produce
ratings, which the sports associations publish; for
example, the Fédération Internationale de Football
Association (FIFA) publishes current FIFA ratings
on its website (http://www.fifa.com/worldfootball/
ranking/index.html). The International Cricket Coun-
cil (ICC) publishes world rankings for national
cricket teams for various formats of the game
on its website (http://www.cricinfo.com/rankings/
content/current/page/211271.html). Although these
systems are referred to as ranking lists, they pro-
vide ratings that somewhat represent team strengths.
Whether official rankings provide accurate estimates
of team strength is an issue that has been addressed in
various sports. The issue is not trivial, and the impor-
tance of such rating systems should not be under-
stated. For example, the FIFA world rankings form
part of the criteria for awarding work permits for
players outside the European Union in the English
Premier League (PL). McHale and Davies (2007) show
that when it comes to forecasting match outcomes,
the world rankings do not perform well. In other
sports, tournament organisers use the official rank-
ings for seeding (e.g., the US Open in tennis), so that
the rankings can influence the tournament outcome.
They might even use rankings to determine qualifi-
cations for a tournament; the top ten teams qualify
automatically for the ICC Cricket World Cup, whereas
lower-ranked teams must play a qualifying event—
the ICC Trophy.
Rating individuals in team sports is a more complex
task, primarily because of the team structure. Players
McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
may play multiple positions. For example, baseball
has specialist batters, pitchers, and fielders, and rat-
ing systems have been developed for each specialty,
as Albert and Bennett (2003), Dewan (2006), and
Keri (2006) discuss. In principle, comparing players
with the same specialty is relatively straightforward,
although several metrics have been suggested for this
purpose. Albert (2009) provides a recent review of
these metrics. However, complications arise when one
wants to compare the contribution (to a team) of a
player with a specialised role with another player
who has a different specialised role. For example,
comparing a specific batter’s performance to that of
a specific pitcher is not a simple task. Goldman and
Kahrl (2010) and James and Henzler (2002) discuss
metrics that attempt such an analysis in baseball—to
value players on a team regardless of their playing
position. Cricket is similar in principle; researchers
have explored methodologies for a common ratings
system (Lewis 2005), as they have with baseball.
Soccer is more complex. Players have particu-
larly varied roles, including goalkeepers, defenders,
midfielders, and attackers (or strikers); each role
requires a different sets of skills. However, each group
also has subgroups. For example, midfielders can
be disaggregated into wingers, defensive midfielders,
and attacking midfielders. A team might not choose
to play players at each position or role, depend-
ing on the strategy and composition of the opposing
team. In addition to the complexity in team makeup,
the interactions between the individuals being rated
and their team members, and the interactions with
the opposing players in the course of a game, make
performance extremely dependent on external fac-
tors that the individual cannot control. For example,
defenders and the goalkeeper (who is the last line of
defence for stopping the opposition from getting the
ball into the goal) who play on winning teams may
have quiet games because their midfield and attack
have been potent. In addition to these complex inter-
actions, players’ roles might change during a game;
for example, fullbacks can play in both attacking and
defensive roles, and would expect to vary their role
depending on the state of the match, the tactics of the
opposition, and instructions from the coach. In addi-
tion, some players might play a variety of positions
during a game; therefore, rating a player who started
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
a game as a midfielder but finished it as a defender
is difficult.
Despite these difficulties, the potential usefulness
of a single rating system is clear. Teams can improve
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
performance by identifying and adding undervalued
players to their rosters, and the popularity of the sport
can be enhanced by increasing debate among fans and
pundits. In recent years, a wealth of data has become
available, making it more attainable to rate all players
using one scale. A rating system or competitor index
could also be considered as a commercial product,
and the information therein can be sold to the media.
In this paper, we describe the development of such
a system, the EA Sports Player Performance Index
(PPI), the official player rating system of the PL (the
top tier of soccer in England) and the Championship
(the second tier of soccer in England). We describe
the structure and construction of the system. We also
highlight issues we encountered during the various
stages of the consultation process and development,
from initial contact to launch, and the challenges of
balancing academic research with industry and cor-
porate requirements. We then consider the values for
players at the time of this writing, and conclude with
a discussion that highlights opportunities for further
development.
Description of the Development Team
for the Index
The EA Sports PPI is a rating system for soccer play-
ers in England. The index is a partnership between
the PL, the Football League (FL), Football DataCo
(FDC), and the Press Association (PA). PL manages
the top tier of soccer in England. FL manages the
second tier of professional soccer in England and for
certain Welsh teams playing in the English leagues.
FDC is a joint venture company owned by the two
professional leagues, which own the rights to data
generated by the professional soccer competitions in
England. PA is the official news agency for the United
Kingdom. EA Sports PPI is the brand name for the
index, which was known as the Actim index until
2010. The index forms part of a wider package of
sports news information that is sold commercially
around the world by PA acting as an agent for FDC.
The development of the index began in 2003. Follow-
ing an initial tender from interested academics, the
341
academic authors of this paper were fortunate to be
given the opportunity to present ideas on how to con-
struct an index for the clients (i.e., PL, FL, FDC, and
PA), and were commissioned as consultants in 2004
to develop the player rating system.
Prior to formal meetings, the clients made it clear
that the new index had to be based on sound math-
ematical reasoning; the media would use it to raise
the Actim brand’s profile. The clients and consultants
agreed that the new index should: (1) be built on
a statistical basis so that subjective opinions are not
used when ranking players or judging the value of
player activity (i.e., judging which players are more
valuable); (2) be able to compare players from dif-
ferent positions; (3) be transparent so that nonstatis-
ticians could understand how points are awarded;
and (4) use a scale with which the general pub-
lic would be comfortable—whole numbers for index
scores were considered desirable. Furthermore, the
clients required that the goals scored should be a
direct component of the final index; they viewed the
number of goals scored as too important a metric to
be omitted from the index, although the conversion
of scoring attempts into goals might be considered
as occurring by chance (Pollard et al. 1977). After an
initial one-year run, the consultants agreed to update
the index to meet additional requirements to include
points for players whose team did not concede a goal
during a match (i.e., a clean sheet) and points for
the player who makes the final pass to a player who
scores a goal (i.e., an assist).
We should explain the rules of the game for the
North American audience. They are essentially very
simple: playing on a pitch of approximately 100 yards
in length, two teams of 11 individuals each attempt
to use their feet and heads (but not hands) to put the
ball into the opponent’s goal; after normal time, the
team with the most goals wins; tournament rules can
vary; some allow extra time to be played to determine
winners and others allow ties (draws).
We also note that the relationship between academic
consultants and industry clients is interesting. Various
authors have examined the academic-industry consul-
tation process, recently and in the past, and in both
operations research (Dudley 1962, Ormerod 2002)
and statistics (Marquardt 1979, Kenett and Thyregod
2006). During the consultancy project that we describe

342
in this paper, the most challenging aspects of the pro-
cess for the academics were managing expectations,
determining desirable properties of the developing
index, and communicating the statistical nuances of
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
the index (e.g., short-term variations). Where the con-
sultant might like to be given carte blanche for the
development of the solution, the client(s) might partly
insist a priori on the form of the solution. The use of
metrics for the number of goals scored, assists, and
clean sheets was an example. The converse of this
point is the acknowledgment that the client is perhaps
better placed to consider the reactions of users of the
solution and the nature of the solution. A compromise
must ultimately be sought between complexity and
simplicity, between the pure and the practical, and
between the radical and the conservative. We elabo-
rate on these experiences later in the text.
The Index
During the period of development, discussion, and
evolution, all parties agreed that the final index would
be a combination of individual component indices,
each with its own properties, construction method,
and emphasis. The individual subindices capture:
(1) match contributions, (2) winning performance,
(3) match appearances, (4) goals scored, (5) assists,
and (6) clean sheets. As mathematical modellers,
index 1 (i.e., match contributions) was of particular
interest to us. However, as a result of high variabil-
ity from match to match and from week to week,
the clients felt that this needed stabilising using more
standard (and simple) measures of performance.
We built the component indices using two sea-
sons of PL data (2002–2003 and 2003–2004). The
PA collected these data, and continue to do so,
using a system called Football Live. This system pro-
vides real-time information on player actions and
match events, essentially to provide ball-by-ball or
kick-by-kick match reports on the Internet and on
mobile platforms. A match reporter (formerly a pro-
fessional player) at the match and a data logger
at the PA base record the Football Live data; the
reporter and data logger are in constant communica-
tion throughout a match via mobile phone link. Some
match action or events (most notably passes and tack-
les) occur too quickly and at too great a frequency to
McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
be captured effectively by the Football Live system.
Postmatch video analysis captures these events. This
data-capture process describes match events (i.e., con-
tributions) and associates a named player with each
event or contribution. Thus, the data contain informa-
tion on match histories of player contributions. The
first component index uses these data to the largest
extent. We call this component index subindex 1 and
describe its concept in the following subsection, leav-
ing the mathematical exposition of its construction
for the appendix. Descriptions of the other subindices
then follow. The final ratings system consists of a
weighted sum of six subindices.
Subindex 1: Modelling Match Outcome
McHale and Scarf (2007) describe a model that relates
shots to player actions. The first subindex of the EA
Sports PPI is based on a simpler version of this model.
The flowchart in Figure 1 represents the levels that
make up this index (considered from the perspec-
tive of the home team). We describe the model in the
following paragraphs and the mathematical develop-
ment of the subindex in the appendix.
We can think of the model in Figure 1, from right
to left of the flow diagram, as follows. The result in
any one game is a function of the goals for and goals
against each team in the game. This function is deter-
ministic; that is, if goals for and goals against are
known, then the game result and hence the points
awarded to each team (three points for a win, one
for a draw (tie), and zero for a loss) are completely
determined.
The goals for and goals against depend on the num-
ber of shots and the shot effectiveness of each team
in the game. We interpret the second stage as a state-
ment of probability, in line with Pollard et al. (1977),
who suggest that goals can be interpreted as a ran-
dom consequence of shots. In our model, the number
of goals by each team is simply the number of shots
each team has, multiplied by the probability that the
shot results in a goal. We call this latter probability
the shot effectiveness; it is the product of the attacking
team’s shot accuracy, the defending team’s blocking
capability, and the defending team’s goalkeeping abil-
ity. The attacking team’s shot accuracy is the probabil-
ity that the home-team shot is on target; the defending
team’s blocking capability is the probability that the
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
Passes, tackles,
crosses, dribbles, Number of
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
blocks, clearances, shots and shot
yellow and red effectiveness
cards
Figure 1: The flowchart depicts an influence diagram for our match outcome model (subindex 1): match result is
a function of goals for and goals against that are determined by the number of shots and shot effectiveness of
the two teams, which are determined by the actions of each team (e.g., passes and tackles).
away team blocks the home-team’s shot conditional
on it being on target; the defending team’s goalkeep-
ing ability is the probability that the away team saves
the shot conditional on it being on target and not
being blocked. This aspect of the model captures the
effect of player contributions relating to shots (on and
off target), blocks, and saves.
We model the number of shots for each team as
independent Poisson random variables with means
that we call the shot rates. The shot rate for a team
is the expected number of shots for that team. We
model this shot rate as a linear function of the number
of player contributions not yet captured in the shot
model (e.g., passes, crosses, dribbles, interceptions,
and red and yellow cards). The linear form allows us
to interpret the coefficient of a player contribution as
the expected number of extra shots when the number
of that contribution increases by one. We treat tack-
les differently in that we consider the ratio of tackles
won to tackles lost as the relevant covariate; treating
tackles won in a linear manner resulted in a nega-
tive coefficient; hence, a player would lose points for
a successful tackle.
Given that the number of shots for the home team
and against the home team are Poisson distributed
and the shot effectiveness is the probability that a shot
results in a goal, and assuming (as we do) that the
shot effectiveness does not vary with the number of
shots, the number of goals for and goals against the
home team are also Poisson-distributed random vari-
ables. The shot process is effectively thinned to give
the goal process. Win and draw probabilities for each
team are then determined using the Poisson probabil-
ities for goals for and goals against (e.g., prob(home
win) = prob411 05 + prob421 05 + prob421 15 + · · · in an
343
Goals for
and Match result
goals against
obvious notation). League points for a team result-
ing from the match outcome then follow; we award
to a player for a single contribution X, the change in
the expected number of points for his team when the
number of contributions X changes by one: that is,
the derivative of the expected number of points with
respect to the number of contributions X. The deriva-
tive for each contribution can be determined from the
double Poisson model for goals for and goals against.
The derivative in each case is evaluated with other
contribution types held at specific values. For con-
tributions in the shot-rate part of the model, these
are team-within-match specific values. This implies
that for these contributions, the points awarded for a
single contribution in the match do not vary within
a team, but do vary between matches and between
teams within a match. For contributions in the shot-
effectiveness part of the model, specific values could
be defined in a number of ways, as we discuss below.
We estimate the exact nature of the shot model from
the data on past matches (i.e., 760 matches in 2002–
2003 and 2003–2004 seasons). As one would expect,
a relationship exists between the number of shots by
a team and the player contributions. Indeed, shots
increase as passes increase. Similarly, as red cards for
the opposition increase, shots are expected to increase.
These assertions are not altogether trivial. For exam-
ple, is it the case that by passing more, you will cre-
ate more shots on goal? Or is the relationship borne
because the better teams, who have more shots, tend
to pass the ball more? We do not attempt to establish
cause and effect, but instead quantify the association
between passes and shots.
As McHale and Scarf (2007) discuss, the relation-
ship between the expected number of shots and
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
344 Interfaces 42(4), pp. 339–351, © 2012 INFORMS

player actions is considered using a regression model.
However, for ease of interpretation of the coefficients
in the model, we use a simplification. We estimate
the model for the expected number of shots by a
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
important contribution to a team’s shot rate is a red
card (i.e., sending off) for an opposition player. Again,
this is not surprising because teams that are down
by one player tend to be dominated. However, red

team using ordinary least squares (not distinguish- cards are not a major overall contributor to the index
ing between the home team and the away team). One because they are relatively rare in comparison to other
could (and probably should) estimate the regression contributions. The signs of the coefficients are all as
model allowing exactly for the Poisson nature of the we would have expected. Where the evidence for a
dependent variable, and using a log-linear link func- real covariate effect is weaker (i.e., larger p value)
tion so that the logarithm of the expected number of for red and yellow cards and clearances, a possible
shots is linear in the player contributions. Additional explanation is that these events tend to occur less
complexity can be added to the modelling by allow- frequently than others. A possible explanation of the
ing for the dependence between the goals scores by weakness of the tackle ratio effect is that the cur-
the two teams in a match. However, our experiments rent model does not ideally capture the influence of
revealed that the final index was no more intuitive tackles. A large number of tackles won, for exam-
with more complicated models for shots. A further ple, can indicate both a team being dominant (win-
advantage of not distinguishing between the home ning the ball back) and being dominated (desper-
and away teams in this context is that it ensures that ate defending). Therefore, the effect of tackles is not
players are awarded the same number of points for straightforward.
actions, regardless of whether they were playing for Estimation of the probabilities in the shot-effective-
the home or away team—a property deemed desir- ness part of the model can proceed in a number of
able by the clients. Table 1 shows the fitted model for ways. The first is to use game-by-game estimates. A
expected number of shots. consequence of this is that the shot effectiveness will
The positive (and statistically significant) coefficient vary from team to team and game to game; a team
on passes, for example, confirms that teams that pass that scores no goals in a game will get no points
more tend to take more shots. Crosses have a large for contributions in that game. The second is to use
coefficient; this is not surprising because a cross into season-long home and away estimates of shot effec-
the opposition penalty area often leads to a goal- tiveness. In this second approach, a home-team con-
scoring opportunity. The largest coefficient here is for tribution will receive the same points regardless of
red cards; hence, our model implies that the most the team and game. Other approaches would be to
estimate shot-effectiveness probabilities by team and
home-away status, but on a season-long basis, so
Variables Coef. Std. err. t-stat p-value
that teams who consistently score more goals would
get more points for contributions such as passes.
Crosses 00519 00069 70490 00000
Dribbles 00118 00026 40600 00000 We could ignore the home-away effect. For the final
Passes 00034 00004 70830 00000 subindex, the clients deemed that it would be most
Opposition interceptions −00024 00009 −20520 00012 desirable to hold shot-effectiveness probabilities con-
Opposition yellows 00253 00134 10890 00059
Opposition reds 10023 00506 20020 00043
stant. This decision was taken partly based on the ide-
Opposition tackle win ratio −00170 00090 −10890 00059 alistic principle that the points awarded to a player
Opposition cleared −00017 00009 −10920 00055 should not depend on the shot effectiveness of other
Constant 60463 00815 70930 00000
players in the game, and partly in response to the
Table 1: The table depicts the results for a model of the shot rate as used higher variability in points awarded to a player for
in our index for rating player performance in a soccer match. For each his actions in a single game using the other proce-
contribution type, the coefficient, its standard error, and t statistic and dures for estimating the shot effectiveness. We also
p-value for test of significance of hypothesis, H0 : coefficient equals zero
are shown. Increasing the number of crosses and increasing the opposi-
considered that players lose points for shots off tar-
tion’s number of red cards are two ways in which a team can increase its get because this may seem harsh on strikers who are
expected number of shots. sometimes given the responsibility of shooting from
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
difficult positions, inducing shots off target. However,
because goals scored contributes to the overall index
(see below), we retained the shots-on, shots-off bal-
ance in subindex 1 in the final implementation.
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
We note that we rescaled the points for saves to
account for two features. First, a large proportion
of saves can be regarded as routine saves. Second,
only one player on the pitch can make a save. This
is in contrast to any other action for which points
are awarded (e.g., tackles or passes for which any
player on the pitch can gain points). The Football Live
data contained information on whether each save was
identified as easy, normal, or difficult. We scale the
points for saves by the proportion of saves that were
identified as difficult. Alternatively, we might have
awarded negative points to goalkeepers for goals
conceded.
Although the model description above is from the
point of view of the home side, away-team points for
players may be calculated by determining the effect
of away-team contributions on the expected number
of points for the away team in a similar manner.
Finally, we rescale the points awarded for each con-
tribution so that the total points over all players and
games in a season is equal to the total league points
in the final table (summed over teams). The points
awarded to a player in a match I1 is the sum over
all contribution types of the product of the number of
contributions by the player multiplied by the value of
each contribution; this value is the expected change
in league points for the team, given one additional
contribution.
Other Subindices
Subindex 2: Points-Sharing Index
For each game, the construction of this subindex uses
the number of minutes played by the player, the
total number of minutes played by all players on his
team, and the number of points the team won in that
game. The points awarded to a player in a match on
subindex 2, I2 , is given by the product of the points
won by the team and the ratio of the number of min-
utes played by the player to the total number of min-
utes played by the entire team, including substitutes.
This subindex effectively shares out the points won
345
by the team among the players who played accord-
ing to how long they spent on the pitch. Note that
the total points awarded over the season (summed
over the players) will equal the total league points in
the final league table (both overall and by the team).
Winning players will lead this subindex; it will have
a mix of strikers, midfielders, defenders, and goal-
keepers because only time on the pitch and results
are considered. Points allocated to players over a sin-
gle weekend are highly variable; one reason is that
many players on losing sides will receive zero points,
perhaps in spite of good individual performances.
Subindex 3: Appearance Index
This subindex divides the number of points won by
all teams in the league among the players accord-
ing to how many minutes they played. This subindex
is similar to subindex 2; however, this index does
not consider results—points are awarded for playing.
Subindex 3 rewards players who play a high pro-
portion of games during the season (i.e., team stal-
warts). The average number of points won in any
one game by any one team is 1.34 (calculated on
the basis of data from the past 10 seasons); thus, the
points awarded to a player in a match on subindex 3,
I3 , is given by the ratio of the number of minutes
played by the player to the total number of minutes
played by the entire team, including substitutes, mul-
tiplied by 1.34. The factor 1.34 is included here so that
the points for this index are on the same scale as the
points for subindices 1 and 2. For subindex 3, it fol-
lows that the total points over all players and games
in a season is equal to the expected total league points
in the final table (summed over teams).
The concept of rewarding players for appearances
is not trivial. For example, teams play more high-
profile games than ever before; as such, the physi-
cal demands placed on players are extremely high.
A common tactic used by coaches is to rest star play-
ers for certain games—either less important games or
perhaps games against lesser opposition. This squad-
rotation policy is a complication for the index because
the best players are not always chosen when they are
available; therefore, the index may not identify the
true quality of the top players because their perfor-
mance is not measured against the full spectrum of
opposition ability.

346
Subindex 4: Goal-Scoring Index
This subindex awards points to players based on
their goals scored. The specific number of points
awarded has been calculated by converting goals into
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
points. At the time of the analysis, over the previ-
ous 12 seasons, 12,492 goals were scored in the top
division and 12,985 league points were won. There-
fore, we estimate how many points one goal is worth
as 121985/121482 = 10039 league points. Thus, on this
index, a player receives 1.039 points for each goal the
player scores. The points awarded to a player in a
match for goals, I4 , is then simply the points per goal
multiplied by the number of goals.
Subindex 5: Assists Index
Analogous to the goals-scored index, each player who
provides an assist gets 1.039 points for that assist.
This puts subindex 5 on the same scale as the other
subindices. The points awarded to a player in a match
for assists, I5 , is then the points per assist multiplied
by the number of assists.
Subindex 6: Clean-Sheets Index
To maintain the balance of the overall index, we take
the total points awarded for clean sheets to be equal
to the total points for assists. In the seasons analysed
(i.e., 760 games in 2002–2003 and 2003–2004), there
were 1,069 assists and 399 clean sheets. Therefore,
we calculate the points per clean sheet as 410039 ×
110695/399 = 20784. We now need to divide these
points among the entire team that achieved the clean
sheet; thus, the points for the clean sheet are not
awarded to a single player. Note that the total points
for a clean sheet is close to, but not quite, three; this
reflects that a clean sheet implies a win—but for a 0–0
result.
The data collected by PA Sport include several
defensive actions: blocks, clearances, tackles won,
interceptions, and saves. A weighted sum of these
actions by position provides a distribution of defen-
sive actions across a team (see Table 2). Thus, fol-
lowing this distribution for the points for a clean
sheet, of the 2.784 points available, a goalkeeper
should get 0021 × 20784 = 00585; each defender, mid-
fielder, and striker should receive a similar number
of points. Thus, if a team achieves a clean sheet, the
subindex points awarded to a player, I6 , are 0.585
McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
Percent of total
Position defensive actions (%)
Goalkeeper 21
Defender (per defender) 13
Midfielder (per midfielder) 5
Striker (per striker) 3
Table 2: In a soccer match, some player actions can be described as
defensive (i.e., blocks, clearances, tackles won, interceptions, and
saves). The table shows the percentage of total defensive actions made
per position (based on four defenders, four midfielders, and two strikers
per team); we calculated the values from the 2002–2003 and 2003–2004
seasons and used them to distribute the points for a clean sheet among
the team.
(goalkeeper), 0.364 (defender), 0.150 (midfielder), and
0.071 (striker). Player positions are defined as at sea-
son start (although player position defined for each
match would facilitate a better assignment of clean-
sheet points).
Final Index
The final index is a weighted sum of the points
achieved on each subindex:
I = 100 × 0025I1 + 00375I2 + 00125I3 + 00125I4


+ 000625I5 + 000625I6 0
A pure index would comprise only subindex 1 or
a variation of it. However, this index does not pro-
vide a practical solution to the problem. Essentially, it
has too much noise to allow its use for reporting on
a weekly basis. Also, subindex 1 is perhaps too radi-
cal given the perception among pundits and fans that
winning (and hence goals, assists, and clean sheets)
determines the best players. Part of this perception
has become firmly established because of the use of
such metrics in fantasy soccer games. The Telegraph
fantasy football website (http://www.fantasyfootball
.telegraph.co.uk) provides details about fantasy soc-
cer. Also, the role of chance in the game is perhaps
too strong in subindex 1. Although uncertainty of out-
come is an important factor in maintaining interest
in the game, observers of the game believe that sys-
tematic effects are more dominant than they are. For
example, a manager who loses x games in a row is
blamed and replaced (Audas et al. 1999, De Schryver
and Eisinga 2009). Thus, the final index comprises
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS 347

a weighted sum of the six subindices. We experi- Subindex 1

Name Team Position points
mented with weighting combinations to find the com-
bination that most satisfied the client’s desire to have
Mark Schwarzer Fulham Goalkeeper 7.29
an index that produced stable ratings, but that was Gareth Barry Aston Villa Midfielder 7.06
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.

varied enough from week to week to generate dis- Sol Campbell Portsmouth Defender 6.86
Gary Cahill Bolton Wanderers Defender 6.70
cussion among fans, the media, and pundits. In this J. Lloyd Samuel Bolton Wanderers Defender 6.63
way, subindices 2–6 can be thought of as stabiliz- Robert Green West Ham United Goalkeeper 6.45
Heurelho Gomes Tottenham Hotspur Goalkeeper 6.43
ing components that regulate the variability inher- David James Portsmouth Goalkeeper 6.40
ent in subindex 1. Thus, the choice of weighting Scott Carson West Bromwich Albion Goalkeeper 6.39
for the subindices reflects the compromise between Nicolas Anelka Chelsea Striker 6.33
Richard Dunne Manchester City Defender 6.30
purity and practicality, and between revolution and Andrew O’Brien Bolton Wanderers Defender 6.26
conformity. Jussi Jaaskelainen Bolton Wanderers Goalkeeper 6.20
Sylvain Distin Portsmouth Defender 6.12
Each subindex possessed the desirable property Thomas Sorensen Stoke City Goalkeeper 6.10
that total points available were approximately equal Matthew Upson West Ham United Defender 6.03
Ryan Shawcross Stoke City Defender 5.85
to league points available in the season; therefore, a
Danny Collins Sunderland Defender 5.81
player’s points might be interpreted as his share of Michael Turner Hull City Defender 5.54
the points won by the team. Despite this property, we Anton Ferdinand Sunderland Defender 5.47
rescaled the final index; when the clients considered
Table 3: For our player rating index based on a model of match outcome
the likely audience for the final index, they decided (subindex 1), the table shows the top 20 players for the 2008–2009 sea-
that it was undesirable to report rating points to two son. The list is dominated by goalkeepers and defenders. The shortage
of strikers is because of the negative impact on subindex 1 of off-target
decimal places. The index is intended to be a talking
shots.
point among fans, which can drive theoretical pro-
jections such as the components of a super team of top teams. One might consider this as having been
top index players in each playing position. Generally, successfully achieved given the appearance of Mark
debate is a large factor in the life of any soccer fan; Schwarzer of Fulham, Joleon Lescott and Tim Howard
such debate takes place in homes, at work, in public of Everton, and Kevin Davies of Bolton Wanderers.
houses, and in cafes, and also on the Internet as fans
embrace the world of blogging and tweeting to settle Name Team Position Index score
the argument of who is the best player in their league
or on their team. Nicolas Anelka Chelsea Striker 751
Dirk Kuyt Liverpool Midfielder 655
Frank Lampard Chelsea Midfielder 653
The Index for the 2008–2009 Season Dimitar Berbatov Manchester United Striker 586
Gabriel Agbonlahor Aston Villa Striker 575
In this section, we present subindex 1 and the final Jose Reina Liverpool Goalkeeper 573
EA Sports PPI (then known as the Actim index) for Gareth Barry Aston Villa Midfielder 564
Jamie Carragher Liverpool Defender 552
the 2008–2009 season of the PL competition. Table 3 Steven Gerrard Liverpool Midfielder 540
shows the subindex 1 final totals. Several observa- Nemanja Vidic Manchester United Defender 540
Cristiano Ronaldo Manchester United Striker 517
tions can be made from the list in Table 3. First, of the Petr Cech Chelsea Goalkeeper 513
top 20 players, 7 are goalkeepers and 11 are defend- John Terry Chelsea Defender 501
Robbie Keane Tottenham Hotspur Striker 482
ers. This is a consequence of the deduction of points
Edwin Van der Sar Manchester United Goalkeeper 481
for off-target shots. The inclusion of goals scored and Neves Denilson Arsenal Midfielder 477
assists in the final index (see Table 4) moderates this Mark Schwarzer Fulham Goalkeeper 477
Joleon Lescott Everton Defender 476
outcome. The final standings in Table 4 have a good Tim Howard Everton Goalkeeper 472
mix of players from different positions: six strikers, Kevin Davies Bolton Wanderers Striker 468
five midfielders, four defenders, and five goalkeepers.
Table 4: The table shows the end-of-season final index totals for the 2008–
One of the original desirable qualities of an index was 2009 season for the top 20 players. The composition is a good mix of
to include players who are not necessarily from the players from all positions and all teams.
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
348 Interfaces 42(4), pp. 339–351, © 2012 INFORMS

8 stable on this subindex. Lampard, infamous for shoot-

Nicolas Anelka
Dirk Kuyt ing off target, is highly variable. The inclusion of
Sub-index 1 cumulative total

6 Frank Lampard the other subindices in the final index moderates this
Mark Schwarzer
Gareth Barry variability.
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.

4 Sol Campbell Following the 2010 World Cup in South Africa,

2 the index with the actual England team chosen is
0 not identical to the team England chose to field in
–2
Aug Sep Oct Nov Dec Jan
Month
Figure 2: The graph shows cumulative totals over time for the top three
players for the 2008–2009 season (Schwarzer, Barry, and Campbell)
according to our subindex 1 and the top three players from the final index
(Anelka, Kuyt, and Lampard). The negative impact of shots off target is
particularly evident for Nicolas Anelka in the early months of the season.
One can see how poor performances result in negative index totals.
Although players in the top teams (i.e., Manchester
United, Liverpool, and Chelsea) lead the index, player
rotation, in which the top teams with large squads
rest their players, means that they do not dominate it.
Figure 2 shows the variability in subindex 1, the
pure contributions index. Although Nicolas Anelka
finished at the top of the overall index, he had
a negative total on subindex 1 for the first two
months of the season. The goalkeeper (Schwarzer)
and defensive midfielder (Barry) are much more
comparing a hypothetical England team picked using
interesting (see Table 5). The team in Table 5 is
the opening game of the World Cup. One omission
is notable: Steven Gerrard of Liverpool is missing.
Feb Mar Apr May Gerrard was injured for a large part of the season;
therefore, he did not receive as many points as he
would have if he had been fit. This raises an impor-
tant point regarding the index—players can only
obtain rating points if they play; therefore, the index
provides a rating of the players who are playing the
best soccer and not necessarily a rating of the best
players. We leave it to the fans to produce a per-
minute-played index.
Discussion
The EA Sports PPI is a performance rating system for
players in the top divisions of soccer in England. It
is now in its seventh season of use. The index was
developed by the authors in collaboration with the
soccer leagues and a leading news media company,
and is a commercial news product that is sold to
news media outlets worldwide. It is published weekly
by the PL. The index is also used as part of the PL

Actim Index ranking Name Team Position Index score Actual player

3 Wayne Rooney Manchester United Striker 482 Wayne Rooney

12 Jermain Defoe Tottenham Hotspur Striker 422 Emile Heskey
9 Frank Lampard Chelsea Midfielder 433 Frank Lamard
16 James Milner Aston Villa Midfielder 376 James Milner
20 Ashley Young Aston Villa Midfielder 370 Aaron Lennon
21 Gareth Barry Manchester City Midfielder 368 Steven Gerrard
19 John Terry Chelsea Defender 371 John Terry
24 Leighton Baines Everton Defender 357 Glen Johnson
27 Ashley Cole Arsenal Defender 348 Ashley Cole
40 Ryan Shawcross Stoke City Defender 325 Ledley King
44 Joe Hart Birmingham City Goalkeeper 322 Robert Green

Table 5: The table shows a hypothetical England team chosen from the highest-ranked players in their playing
position as of March 12, 2010 (column 2), and the actual team that played the first game in the 2010 World Cup
finals (final column). The two lineups are closely matched, especially when we consider that Steven Gerrard and
Rio Ferdinand were injured for a large part of the season; thus, they were unable to accumulate points.
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
Fantasy Football game, and is published in match-
day programmes and newspapers around the globe.
The index was the first to rate individual players in
a team sport using a single common score, regard-
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
less of a player’s position. This single-rating score
was a particular challenge, which we addressed dur-
ing the index development. We were able to meet
this objective partly because soccer players multitask;
although players have specialities, roles can change
subtly within a game, and all players have the oppor-
tunity to make a range of winning contributions,
for example, passes, tackles, and shots. Goalkeepers
also make contributions other than saves, although
infrequently.
The PL and Championship in England use the in-
dex in their annual competitions, and the PL uses
it in its Fantasy Football competition. The index is
updated weekly during the season; the list of cur-
rent top 100 players in the PL is published on the
PL website (http://www.premierleague.com/page/
ActimIndex) each Tuesday. Lists of best defenders,
midfielders, attackers, and goalkeepers are also pub-
lished, as is the EA sports team of the week. The
team of the week comprises the players who score
mostly highly on the index during the previous week,
subject to the constraint that the team must include
one goalkeeper, four defenders, four midfielders, and
two attackers. The team of the week consequently
changes on a week-by-week basis. The overall index
evolves more slowly, given that a player’s index points
accumulate over the season; an injury or layoff can
lead to a significant reversal of fortune for a player.
The ratings and rankings are also published in news-
papers and match programmes, and are debated by
observers and commentators worldwide. An Actim
index blog (http://www.actimindex.wordpress.com/
the-actim-index) is also published (note that accessing
the blog requires user registration).
The availability of detailed data collected both
during and after matches regarding players’ contri-
butions and the notion that such data capture the
essence of the beautiful game essentially drove the
development of the index. However, we found that
an index based purely on player contributions was
too volatile for weekly publication. The clients for the
index have a more deterministic focus, and the final
349
index comprises a weighted average of those con-
tributions that directly influence success (e.g., league
points won, goals) and those contributions that statis-
tically influence success (e.g., passes, tackles). In this
latter case, we mean those contributions that appear,
on the basis of our statistical analysis, to be associated
with goals scored.
As data relating to player contributions become
more sophisticated and more representative of the
game, we expect further development of the index.
For example, the collection of pass origin and fin-
ish is now feasible, as is the collection of data relat-
ing to sequences of passes. Such data could allow
complex interactions between player actions to be
modelled in a way that is not currently considered
in the index construction. Additionally, for season
2010–2011, the PL is introducing a player tracking
system that will provide distance covered by all mov-
ing objects (including the ball) engaged on the field
of play, coupled with event-time recording to provide
additional calculations relating to events in the game
(e.g., speed, distance, acceleration, deceleration, force
of pass). The future of sports statistics is becoming
ever more sophisticated.
Appendix
In the model development, we use the following notation
for a game.
HP T S : number of points for the home side (0, 1, or 3),
depending on outcome.
AP T S : number of points for the away team.
HG : number of goals for the home team.
AG : number of goals for the away team (goals against the
home team).
HX : number of contributions X by the home team—X
may refer to passes, tackles, etc.
AX : number of contributions X by the away team.
HS 1 AS : number of shots (home, away)
E8HS 9: home-team expected number of shots per game
(i.e., shot rate), respectively.
pHG — HS 1 pAG — AS : probability of a home (away) goal given a
shot by the home (away) team (shot-effectiveness).
pHSAVE — AS : probability of a home save given an away-team
shot.
pHB — AS : probability of a home-team block given an away-
team shot.
pHS_ON — HS : probability of a home-team shot on target
given a home-team shot.
In terms of the three stages discussed following Figure
1, the model structure is as follows. The points for the

350
home side, HP T S , depend on the goals for and goals against,
so that  derivatives, ¡E8HP T S 9/¡HX , can be determined from the
3 if HG > AG
 equation above and Equations (1), (2), (4), and (5). We omit
HP T S = 1 if HG = AG

0 if H < A 1
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
G G
because teams are awarded three points for a win, one point
for a draw, and zero points for a loss. At the next level in
the hierarchy, we model the shot effectiveness and the shot
rate. For the shot effectiveness, we assume
pHG — HS = 41 − pAB — HS 541 − pASAVE — HS 5pHS_ON — HS 1
where pAB — HS is the probability that the away team blocks
the home team’s shot, pASAVE — HS is the probability that the
away team saves the shot, and pHS_ON — HS is the probability
that the home-team shot is on target. For the away team we
have
pAG — AS = 41 − pHB — AS 541 − pHSAVE — AS 5pAS_ON — AS 0
The shot effectiveness is the probability of a goal given
a shot; this is assumed to be the product of the probabil-
ities that the shot is on target (pAS_ON — AS 5, is not blocked
41 − pHB — AS 5, and is not saved 41 − pHSAVE — AS 5.
We model the goals for and goals against as indepen-
dent Poisson-distributed random variables, so that HG ∼
Poisson4pHG — HS E8HS 95, where pHG — HS is the effectiveness of
the home-team (away-team) shots (shot effectiveness) and
E8HS 9 is the expected number of home-team (away-team)
shots per game. We model the expected number of shots
(i.e., the shot rate) as a function of other aspects of play
such as passes, dribbles, etc. Thus,
E8HS 9 = f 4HX 1 HY 1 0 0 0 1 AU 1 AV 1 0 0 051
and similarly for the away team. HX is, for example, the
number of contributions of type X (e.g., passes) by the home
side and AU is the number of contributions of type U (e.g.,
interceptions) by the away side. It is natural to use the
exponential function for the function in Equation (3) above,
although it is simpler to assume that
E8HS 9 = 0 +X HX +Y HY +···+‚U AU +‚V AV +··· 0
With this form, we can interpret the coefficient P , for
example, as the expected number of extra shots when the
number of passes increases by one.
Denoting the expected number of points for the home
side by E8HP T S 9, the number of points we award to a player
for a single contribution X is the change in the expected
number of points for his team when the number of contri-
butions X changes by one, that is, ¡E8HP T S 9/¡HX . Noting
that the double Poisson model implies
E8HPTS 9
= 43 points5×4probability of a win5+41 point5
×4probability of a draw5
 ˆ ˆ j
X X ‹HG ‹iAG X ˆ
‹iHG ‹iAG

= exp4−‹HG −‹AG 5 3 + 1 (5)
i=0 j=i+1
i!j! i=0
i!i!
McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
where ‹HG = pHG — HS E8HS 9 and ‹AG = pAG — AS E8AS 9, all the
these derivatives for brevity, but note that for each contribu-
tion, the derivative is evaluated with all other contribution
types held at their match and team-within-match specific
values. Although the model description above is from the
point of view of the home side, away-team points for play-
ers may be calculated by determining the effect of away-
team contributions on the expected number of points for
(1)
the away team, ¡E8AP T S 9/¡AX , in a similar manner.
Estimation of pHG — HS and pAG — AS can proceed in a num-
ber of ways. The first is to use game-by-game estimates, so
that p̂HG — HS = HG /HS and p̂AG — AS = AG /AS . The second is to
use season-long home and away estimates of shot effective-
ness, that is, p̂HG — HS = H̄G /H̄S and p̂AG — AS = ĀG /ĀS , where H̄G
(2) and H̄S are the season-long arithmetic means of the num-
ber of home goals and home shots per game, respectively
(and similarly for away goals and shots). Another approach
would be to estimate pHG — HS and pAG — AS by team and home-
away, but on a season-long basis. The home-away effect can
be ignored by setting
p̂HG — HS = p̂AG — AS = 4H̄G + ĀG 5/4H̄S + ĀS 50
The points awarded to a home-team player i in a match,
P
I1 , is then X Hi1 X 8¡E4HP T S 5/¡X9, where Hi1 X is the number
of contributions of type X by player i; the formulation for
away-team players is similar.
For subindex 2, the points awarded to player i in match
(3) t is given by
minsit
I2 = P × pointst 1
i minsit
where minsit is the number of minutes player i was on the
P
pitch in game t, i minsit is the total number of minutes
that all players on the team of player i (including player i)
were on the pitch (approximately 90 minutes × 11 players)
(4) in game t, and pointst is the league points won by the team
of player i in match t, being equal to 3, 1, or 0 according to
the outcome of game t.
For subindex 3, the points awarded to player i in match
t are given by
minsit
I3 = P × 10340
i minsit
The other subindices are described in the main text.
Acknowledgments
The authors acknowledge the contribution of several people
in developing the Actim index: Martin Axford (PL), Chris
Buckley (ex PA), Neil Chambers (PA), Andy Elliot (ex PA),
Guy Hollis (PA), Phil Lines (PL), Paul Molnar (PL), and
Brian Phillpotts (ex PL). They also thank two anonymous
referees and the associate editor for their suggestions, which
have improved the paper.
 McHale, Scarf, and Folker: Soccer Player Performance Rating System
Interfaces 42(4), pp. 339–351, © 2012 INFORMS
References
Albert, J. 2009. Exploring pitch F/X data. R. Koning, P. Scarf, eds.
Proc. 2nd Internat. Conf. Math. Sport, the Institute of Mathemat-
ics and Its Applications, London, UK, 10–18.
Downloaded from informs.org by [157.182.150.22] on 08 June 2014, at 12:47 . For personal use only, all rights reserved.
Albert, J., J. Bennett. 2003. Curve Ball. Springer, New York.
Audas, R., S. Dobson, J. Goddard. 1999. Organizational perfor-
mance and managerial turnover. Managerial Decision Econom.
20(6) 305–318.
De Schryver, T., R. N. Eisinga. 2009. The application of piecewise
linear models to management turnover decisions in Dutch soc-
cer clubs. R. Koning, P. Scarf, eds. Proc. 2nd Internat. Conf.
Math. Sport, the Institute of Mathematics and Its Applications,
London, 159–164.
Dewan, J. 2006. The Fielding Bible. ACTA Sports, Chicago.
Dudley, N. A. 1962. Operational research and the universities.
J. Oper. Res. Soc. 13(1) 81–86.
Goldman, S., C. Kahrl. 2010. Baseball Prospectus 2010: The Essen-
tial Guide to the 2010 Baseball Season. John Wiley & Sons, Inc.,
Hoboken, NJ.
Grinblatt, M., S. Titman. 1989. Portfolio performance evaluation:
Old issues and new insights. Rev. Financial Stud. 2(3) 393–421.
James, B., J. Henzler. 2002. Win Shares. STATS Publishing Inc.,
Northbrook, IL.
Kenett, R., P. Thyregod. 2006. Aspects of statistical consult-
ing not taught by academia. Statistica Neerlandica 60(3)
396–411.
Keri, J. 2006. Baseball Between the Numbers. Basic Books, New York.
351
Lewis, A. J. 2005. Towards fairer measures of player performance
in one-day cricket. J. Oper. Res. Soc. 56(7) 804–815.
Lewis, M. M. 2003. Moneyball: The Art of Winning an Unfair Game.
W. W. Norton and Co., New York.
Marquardt, D. W. 1979. Statistical consulting in industry. Amer.
Statistician 33(3) 102–107.
McHale, I. G., S. M. Davies. 2007. Statistical analysis of the FIFA
world rankings. R. Koning, J. Albert, eds. Statistical Thinking in
Sport. Chapman and Hall, Boca Raton, FL, 77–90.
McHale, I. G., D. Forrest. 2005. The importance of recent scores in
a forecasting model for professional golf tournaments. IMA J.
Management Math. 16(2) 131–140.
McHale, I. G., A. Morton. 2011. A Bradley-Terry type model for
forecasting tennis match results. Internat. J. Forecasting 27(2)
619–630.
McHale, I. G., P. A. Scarf. 2007. Modelling soccer matches using
bivariate discrete distributions with general dependence struc-
ture. Statistica Neerlandica 61(4) 432–445.
Ormerod, R. J. 2002. On the nature of OR: Taking stock. J. Oper. Res.
Soc. 53(5) 475–491.
Pollard, R., B. Benjamin, C. Reep. 1977. Sport and the negative
binomial distribution. S. P. Ladany, R. E. Machol, eds. Optimal
Strategies in Sports. North Holland, New York, 118–195.
Swamidass, P. M., W. T. Newell. 1987. Manufacturing strategy,
environmental uncertainty and performance: A path analytic
model. Management Sci. 33(4) 509–524.